CONN
Default preprocessing pipeline
Functional realignment and unwarp
Scans are coregistered and resampled into a reference image (first scan of session) with b-spline interpolation.
Addresses potential susceptibility distortion-by-motion interactions by estimating derivatives of deformation field with respect to head movement, then resamples functional data to match the deformation field of the reference image.
If double-echo sequence is available, field inhomogeneity inside the scanner (fieldmap) is estimated and used for Susceptibility Distortion Correction (SDC) as part of the unwarp step → functional data is resampled along phase-encoded direction in order to correct absolute deformation state of reference image.
Slice-timing correction
Temporal misalignment between slices of functional data is introduced by sequential nature of fMRI acquisition protocol.
Corrected with SPM12 STC procedure where functional data is time-shifted and resampled with sinc-interpolation to match time in middle of each acquisition time (TA)
Outlier identification
Outlier scans are detected from observed global BOLD signal and amount of subject-motion in the scanner.
If an acquisition has a framewise displacement above 0.9mm or global BOLD signal changes about 5 s.d., it is flagged as a potential outlier. (these parameters can be changed for more conservative thresholds).
Framewise displacement is computed at each timepoint by considering a 140x180x115mm bounding box around the brain and estimating the largest idsplacement among six control points placed at the center of the bounding-box faces.
Global BOLD signal change is computed at each timepoint as the change in average BOLD signal within SPM’s global-mean mask scaled ot s.d. units
Direct segmentation and normalization
Anatomical and functional data are normalized into standard MNI space and segmented into GM, WM, CSF tissue classes with SPM12 unified segmentation and normalization procedure.
Performs iterative tissue classification by estimating posterior tissue probability maps (TPMs) from intensity values of reference functional/anatomical image.
Registration estimates the non-linear spatial transformation which best approximates posterior and prior TPMs until convergence.
Direct normalization applies unified segmentation and normalization procedure separately to functional data using mean BOLD signal as reference image, and to structual data using raw T1-weighted volume as reference.
Functional and anatomical data are resampled to default 180x216x180mm bounding box with 2mm isotopic voxels for functional data and 1mm for anatomical using 4th order spline interpolation.
Direct vs. indirect
Direct normalization → functional and anatomical data are normalized separately using their own nonlinear transformations to project data into MNI space.
Indirect → both sets use the same nonlinear transformation estimated using only the structural data
Functional smoothing
Functional data is smoothed with spatial convolution with a Gaussian kernel of xmm full width half maximum (FWHM) to increase BOLD signal:noise ratio and reduce influence of residual variability in functional and gyral anatomy across subjects
Indirect segmentation and normalization
Functional data is coregistered with an affine transformation to structural data with SPM12 intermodality coregistration procedure with normalized mutual information cost function.
Then anatomical data is normalized into standard MNI space and segmented (GM, WM, CSF) with SPM12 unified segmentation & normalization procedure. (same as direct). This procedure is applied to structural data using the raw T1-weighted volume as a reference, and then applies the same estimated transformation to the functional data.
Functional/anatomical Coregistration
Functional data is co-registered to the structural data using using SPM12 inter-modality coregistration procedure with a normalized mutual information cost function (Collignon et al. 1995, Studholme et al. 1998). This procedure estimates an optimal affine transformation between the reference functional image (mean BOLD signal) and the reference structural image (T1-weighted volume) that maximizes the mutual information between the two, storing this information in the functional image voxel-to-world mapping header information without resampling the data
Default denoising pipeline
Linear regression
Factors identified as potential confounding effects to estimated BOLD signal are estimated and removed separately for each voxel and for each subject & functional run/session with Ordinary Least Squares (OLS) regression, which projects each BOLD signal timeseries to subspace orthogonal to all potential confounding effects.
These potential confounding effects implement anatomical component-based noise correction procedure (aCompCor) instead of Global Signal Regression (GSR).
- Includes noise components from cerebral white matter and cerebrospinal areas, estimated subject-motion parameters, outlier scans and scrubbing, constant & first-order linear session effects, constant task effects
Temporal band-pass filtering
Temporal frequencies below 0.008Hz or above 0.09Hz are removed to focus on slow-frequency fluctuations while minimizing influence of physioloigcal, head-motion and other noise sources.
Filtering is performed with a discrete cosine transform windowing operation to minimize border effects
Evaluating denoising outputs
The effect of denoising can be seen by estimating distribution of FC values b/w randomly-selected pairs of points within the brain before and after denoising.
Before denoising, FC distributions exhibit large intersession & intersubject variability, as well as skewed distributions w/ varying degrees of positive biases consistent with influence of global/large-scale physiological & subject-motion effects.
After denoising, you see approximately centred distributions with small but noticeable tails on the positive side.
You can also compute QC-FC correlations which examine FC values between randomly-selected brain point pairs and correlate them across subjects with other QC measures like subject-motion indicators.
This method computes a QC-FC correlation for each randomly-selected pair of points and displays the distirbution of resulting QC-FC correlation values.
Alternatives & additions for denoising
ICA Denoising (Griffanti et al 2017) → data-driven approach using ICA to identify potential noise-related temporal components
Retroicor (Glover et al 2000) → uses cardiac & respiratory state information to build a series of predicted sine and cosine components of these effects, which can be used as a confounding effect in CONN’s linear regression step
Simult (Hallquist et al 2013) → alternative implementation of standard sequential regression followed by filtering approach. Both regression & filtering are implemented simultaneously as a single regression step. This is equivalent to removing the frequency-specific effect of potential confounder variables.
Friston24 (Friston et al 1996) → richer set of motion-related regressors designed to remove an autoregressive-moving-average model of effects of subject motion.
Global Signal Regression (GSR) → uses average BOLD signal across entire brian as a potential confounding effect. Not generally recommended because it can introduce artifactual biases (Murphy et al 2009) and remove potentially meaningful neural components (Chai et al 2012) as well as introduce confounding effects across populations.
Connectivity measures
Related
References
https://web.conn-toolbox.org/fmri-methods/preprocessing-pipeline